Swin Transformer-based Cuffless Blood Pressure Estimation using Cardiovascular Signals

Abstract

A conventional sphygmomanometer used for blood pressure (BP) measurement is considered as a golden standard. However, this bulky device requires a pump, valve, battery, and relevant components to inflate and deflate the cuff. Prolonged BP monitoring using a sphygmomanometer may create discomfort for the patients. Therefore, cuffless blood pressure estimation has recently emerged as an excellent choice. Cuffless BP estimation can be performed using electrocardiogram (ECG) and photoplethysmogram (PPG) signals or both ECG and PPG measurements. Various machine learning-based techniques are proposed for cuffless BP estimation utilizing different time-domain and frequency-domain features of ECG and PPG signals. However, the accuracy of the measurements is one of the major concerns. BP estimation using the higher-order spectral features, including bispectrum, has yet to be explored in the literature. This paper proposes an effective approach that utilizes the bispectrum features and a Swin transformer-based network. Capturing the spectral variations in non-stationary cardiovascular signals is crucial. This task is accomplished by signal-to-image conversion of the cardiovascular measurements and utilizing the bispectrum features as input to the Swin transformer. The transformer architecture incorporates a layer normalization, two-level multilayer perception, and a shifted window-based multi-head self-attention (SW-MSA) mechanism. The proposed method can handle long-range dependency and learn global structures effectively. The proposed technique is validated by considering various international standards for cuffless BP estimation tasks. A comparative study is performed on different evaluation metrics concerning contemporary techniques. © 2024 IEEE.